Legal claims defining the scope of protection. Each claim is shown in both the original legal language and a plain English translation.
1. A method for detecting a debatable document performed in a computing device comprising one or more processors and a memory storing one or more programs to be executed by the one or more processors, the method comprising: receiving a document comprising one or more sentences; generating an embedding vector for each of words included in the document; and extracting features of the document from an embedding vector matrix comprising the embedding vectors for the words, and detecting debatability of the document from the extracted features through a detection model comprising a first-step convolutional neural network and a second-step convolutional neural network, wherein each of the first-step convolutional neural network and the second-step convolutional neural network comprises a convolution layer.
This invention relates to natural language processing and text analysis, specifically detecting whether a document contains debatable content. The problem addressed is the automated identification of text that may be controversial, subjective, or open to interpretation, which is useful for applications like fact-checking, moderation, or content filtering. The method processes a document by first generating embedding vectors for each word in the document, which represent semantic relationships between words. These vectors are combined into a matrix, from which features are extracted. The extracted features are then analyzed using a two-step convolutional neural network (CNN) architecture. The first CNN layer processes the features at a lower level, while the second CNN layer refines the analysis at a higher level. The output of the second CNN determines whether the document is debatable. The use of two CNN layers allows for hierarchical feature extraction, improving accuracy in identifying debatable content. The method is implemented in a computing device with processors and memory, ensuring efficient and scalable analysis of text documents.
2. The method of claim 1 , wherein the detection model comprises the first-step convolutional neural network comprising a first convolution layer for outputting a first feature vector by performing a convolution operation between the embedding vector matrix and a plurality of filters, and a first pooling layer for outputting a second feature vector by performing sub-sampling to the first feature vector; and the second-step convolutional neural network comprising a second convolution layer for outputting a third feature vector by performing a convolution operation between the second feature vector and a plurality of filters, and a second pooling layer for outputting a fourth feature vector by performing sub-sampling to the third feature vector.
This invention relates to a method for processing data using a multi-step convolutional neural network (CNN) architecture. The method addresses the challenge of efficiently extracting and refining feature representations from input data, particularly in applications like image or signal analysis where hierarchical feature extraction is beneficial. The method involves a two-step CNN structure. In the first step, an input embedding vector matrix undergoes convolution with multiple filters in a first convolution layer, producing a first feature vector. This feature vector is then processed by a first pooling layer, which performs sub-sampling to generate a second feature vector. The second step involves a second convolution layer, where the second feature vector is convolved with another set of filters to produce a third feature vector. This third feature vector is further refined by a second pooling layer, which performs sub-sampling to output a fourth feature vector. The hierarchical design allows for progressive abstraction of features, improving the model's ability to capture complex patterns in the input data. The method is particularly useful in tasks requiring multi-level feature extraction, such as computer vision or natural language processing.
3. The method of claim 2 , wherein the first convolution layer and the second convolution layer perform the convolution operation using a hyperbolic tangent function or a rectified linear unit (ReLU) function as an activation function.
This invention relates to neural network architectures, specifically convolutional neural networks (CNNs) used for image or signal processing tasks. The problem addressed is improving the efficiency and performance of convolutional layers in CNNs by optimizing the activation functions applied during convolution operations. Traditional CNNs often use fixed activation functions, which may not be optimal for all layers or tasks, leading to suboptimal learning and computational inefficiency. The invention describes a method for performing convolution operations in a neural network where at least two convolution layers are involved. The first convolution layer and the second convolution layer apply a convolution operation using either a hyperbolic tangent (tanh) function or a rectified linear unit (ReLU) function as the activation function. The tanh function constrains outputs to a range between -1 and 1, which can help stabilize training, while ReLU allows for faster convergence by eliminating negative activations. The choice of activation function can be tailored to the specific requirements of the layer, such as preserving gradient flow or accelerating computation. This approach enhances the adaptability and performance of the CNN by allowing different activation functions to be applied at different stages of the network, improving accuracy and training efficiency. The method is applicable to various CNN-based applications, including image recognition, object detection, and signal processing.
4. The method of claim 2 , wherein the first pooling layer and the second pooling layer perform the sub-sampling using a max pooling function.
The invention relates to neural network architectures, specifically focusing on convolutional neural networks (CNNs) used for image or data processing tasks. A common challenge in CNNs is efficiently reducing the spatial dimensions of feature maps while preserving important information, typically achieved through pooling layers. Traditional pooling methods, such as average or max pooling, can impact computational efficiency and feature representation. The invention addresses this by implementing a CNN architecture with at least two pooling layers that perform sub-sampling using a max pooling function. Max pooling selects the maximum value within a defined window, which helps retain the most salient features while reducing dimensionality. This approach enhances computational efficiency and improves feature extraction by focusing on the most prominent activations. The pooling layers are applied sequentially, ensuring progressive down-sampling of the input data. The use of max pooling in both layers ensures consistency in feature retention and dimensionality reduction, making the network more robust to variations in input data. This method is particularly useful in applications like image classification, object detection, and other tasks requiring hierarchical feature learning. The invention optimizes the trade-off between computational cost and feature preservation, leading to improved performance in deep learning models.
5. The method of claim 2 , wherein the detection model further comprises: one or more fully-connected layers connected to the second pooling layer; and an output layer for outputting a discrimination value of the debatability of the document from outputs of the one or more fully-connected layers.
This invention relates to a machine learning-based system for detecting the debatability of documents, such as text or other content, by analyzing their features to determine whether the content is likely to be controversial or subject to dispute. The system addresses the challenge of automatically identifying content that may require further review or moderation, particularly in applications like social media, news platforms, or legal document analysis. The method involves processing a document through a neural network architecture designed to assess its debatability. The network includes an input layer that receives document features, followed by a first convolutional layer that extracts local patterns from the input. A first pooling layer then reduces the dimensionality of the extracted features. A second convolutional layer further processes these features, followed by a second pooling layer to refine the representation. The system then applies one or more fully-connected layers to aggregate the pooled features, enabling the network to learn higher-level relationships. Finally, an output layer generates a discrimination value indicating the likelihood that the document is debatable, providing a quantitative measure of its potential controversy. This approach leverages deep learning techniques to automate the assessment of document debatability, improving efficiency in content moderation and analysis tasks. The architecture ensures that both local and global features of the document are considered, enhancing the accuracy of the discrimination value.
6. The method of claim 5 , wherein the output layer outputs the discrimination value using a softmax function as an activation function.
A method for improving classification accuracy in machine learning models involves using a softmax function in the output layer to generate discrimination values. The softmax function converts raw model outputs into probabilities that sum to one, enhancing interpretability and decision-making. This technique is particularly useful in multi-class classification tasks where distinguishing between multiple categories is critical. The method ensures that the output layer produces normalized probabilities, which helps in selecting the most likely class with higher confidence. By applying the softmax function, the model can effectively handle scenarios where multiple classes are closely related or ambiguous, improving overall classification performance. This approach is commonly used in neural networks, particularly in tasks like image recognition, natural language processing, and other domains requiring precise classification. The softmax function's normalization property ensures that the model's predictions are well-calibrated, reducing the risk of overconfident or inconsistent outputs. This method is part of a broader system for training and deploying machine learning models, where the output layer's design plays a crucial role in achieving accurate and reliable results.
7. The method of claim 1 , wherein the generating of the embedding vector comprises converting each of the words included in the document into a one-hot vector, and generating the embedding vector for each of the words by a product of the embedding matrix and the one-hot vector for each of the words.
This invention relates to natural language processing (NLP) and text representation techniques, specifically addressing the challenge of converting textual data into numerical embeddings for machine learning applications. The method involves generating an embedding vector for each word in a document by first converting each word into a one-hot vector, which is a sparse binary vector representing the word's presence in a predefined vocabulary. This one-hot vector is then multiplied by an embedding matrix to produce a dense embedding vector. The embedding matrix is a learned transformation that maps the discrete one-hot representation into a continuous vector space, capturing semantic relationships between words. This approach enables efficient and meaningful text representation, improving tasks such as document classification, semantic search, and machine translation by preserving contextual information in a compact numerical form. The method ensures that words with similar meanings are positioned closer in the embedding space, enhancing the performance of downstream NLP models. The technique is particularly useful in applications requiring high-dimensional text data processing, where traditional bag-of-words models lack the ability to capture semantic nuances.
8. An apparatus for detecting debatable document, the apparatus comprising: one or more hardware processors and one or more computer readable media storing instructions that, when executed by the one or more hardware processors, cause the apparatus to: receive a document comprising one or more sentences; generate an embedding vector for each of words included in the document; and extract features of the document from an embedding vector matrix comprising the embedding vectors for the words, and detect debatability of the document from the extracted features through a detection model comprising a first-step convolutional neural network and a second-step convolutional neural network, wherein each of the first-step convolutional neural network and the second-step convolutional neural network comprises a convolution layer.
This invention relates to a system for detecting debatable content in documents using machine learning techniques. The problem addressed is the automated identification of text that may contain controversial, subjective, or disputed information, which is valuable for applications like fact-checking, content moderation, and misinformation detection. The apparatus includes hardware processors and computer-readable media storing instructions to process a document. The system receives a document containing one or more sentences and generates an embedding vector for each word in the document. These vectors are combined into an embedding vector matrix, which represents the semantic and contextual relationships between words. Features are then extracted from this matrix to analyze the document's content. The detection model uses a two-step convolutional neural network (CNN) architecture. The first-step CNN processes the embedding matrix to capture local patterns and relationships between words. The second-step CNN further refines these features to assess the overall debatability of the document. Each CNN includes a convolution layer to identify relevant patterns in the data. The system outputs a determination of whether the document contains debatable content based on the extracted features and the trained model. This approach leverages deep learning to automate the detection of subjective or controversial text, improving efficiency in content analysis tasks.
9. The apparatus of claim 8 , wherein the detection model comprises: the first-step convolutional neural network comprising a first convolution layer for outputting a first feature vector by performing a convolution operation between the embedding vector matrix and a plurality of filters, and a first pooling layer for outputting a second feature vector by performing sub-sampling to the first feature vector; and the second-step convolutional neural network comprising a second convolution layer for outputting a third feature vector by performing a convolution operation between the second feature vector and a plurality of filters, and a second pooling layer for outputting a fourth feature vector by performing sub-sampling to the third feature vector.
This invention relates to a machine learning apparatus for processing input data using a multi-step convolutional neural network (CNN) architecture. The apparatus addresses the challenge of efficiently extracting hierarchical features from input data, such as text or images, by leveraging a two-stage CNN structure to progressively refine feature representations. The apparatus includes a detection model that processes an embedding vector matrix, which is a numerical representation of the input data. The first-step CNN consists of a convolution layer that applies multiple filters to the embedding vector matrix, producing a first feature vector. This vector undergoes sub-sampling in a pooling layer, generating a second feature vector with reduced dimensionality. The second-step CNN further refines this output by applying another convolution layer to the second feature vector, producing a third feature vector. A second pooling layer then performs sub-sampling on the third feature vector, yielding a fourth feature vector. This hierarchical processing enhances the model's ability to capture both low-level and high-level features, improving accuracy in tasks like classification or detection. The apparatus is particularly useful in applications requiring robust feature extraction from complex input data.
10. The apparatus of claim 9 , wherein the first convolution layer and the second convolution layer perform the convolution operation using a hyperbolic tangent function or a rectified linear unit (ReLU) function as an activation function.
The invention relates to a neural network apparatus for image processing, specifically focusing on convolutional neural networks (CNNs) used for tasks such as image recognition or classification. The problem addressed is improving the efficiency and accuracy of feature extraction in CNNs by optimizing the activation functions used in convolution layers. Traditional activation functions may not effectively capture complex patterns in image data, leading to suboptimal performance. The apparatus includes a first convolution layer and a second convolution layer, each performing convolution operations on input data. The key improvement is the use of specific activation functions—either a hyperbolic tangent (tanh) function or a rectified linear unit (ReLU) function—to enhance the network's ability to learn and represent features. The tanh function maps inputs to a range between -1 and 1, providing smooth gradients for training, while ReLU introduces non-linearity by zeroing out negative values, which can accelerate convergence. These activation functions are applied after the convolution operations to transform the output before passing it to subsequent layers. The apparatus may also include additional components, such as pooling layers for dimensionality reduction or fully connected layers for final classification. The choice of activation function depends on the specific requirements of the task, with tanh being useful for bounded outputs and ReLU for faster training in deep networks. This design aims to improve the network's performance in tasks requiring high accuracy and computational efficiency.
11. The apparatus of claim 9 , wherein the first pooling layer and the second pooling layer perform the sub-sampling using a max pooling function.
This invention relates to a neural network apparatus designed for image processing, specifically focusing on improving feature extraction through pooling operations. The apparatus includes a convolutional neural network (CNN) with multiple layers, including at least two pooling layers that perform sub-sampling to reduce spatial dimensions of feature maps while retaining important information. The key innovation lies in the use of a max pooling function within these pooling layers, which selects the maximum value from each sub-region of the input feature map, enhancing computational efficiency and robustness to variations in input data. The apparatus may also include convolutional layers that apply filters to extract hierarchical features from the input image, followed by activation functions to introduce non-linearity. The pooling layers are strategically placed after convolutional layers to progressively downsample the feature maps, reducing the computational load while preserving critical spatial information. This design is particularly useful in tasks like image classification, object detection, and segmentation, where efficient feature extraction is essential. The max pooling function ensures that the most salient features are retained, improving the network's ability to generalize across different input variations. The apparatus may be implemented in hardware or software, depending on the application requirements, and can be integrated into larger systems for real-time image processing.
12. The apparatus of claim 9 , wherein the detection model further comprises one or more fully-connected layers connected to the second pooling layer; and an output layer for outputting a discrimination value of the debatability of the document from outputs of the one or more fully-connected layers.
This invention relates to a machine learning-based apparatus for assessing the debatability of documents, addressing the challenge of automatically determining whether a document's content is likely to be controversial or subject to disagreement. The apparatus includes a detection model designed to analyze text and output a discrimination value indicating the likelihood of the document being debatable. The model incorporates a convolutional neural network (CNN) with multiple convolutional layers for feature extraction, followed by a pooling layer to reduce dimensionality. A second pooling layer further refines the extracted features. The model includes one or more fully-connected layers connected to the second pooling layer, which aggregate and process the features. An output layer then generates a discrimination value based on the processed features, quantifying the document's debatability. The apparatus may also include a preprocessing module to prepare the input text and a training module to optimize the model using labeled data. The system enables automated assessment of content controversy, useful for applications like moderation, fact-checking, or content recommendation.
13. The apparatus of claim 12 , wherein the output layer outputs the discrimination value using a softmax function as an activation function.
A system for classification tasks in machine learning processes a set of input data through a neural network to generate a discrimination value. The neural network includes multiple layers, including an input layer, one or more hidden layers, and an output layer. The hidden layers apply nonlinear transformations to the input data, and the output layer produces a final output. The output layer uses a softmax function as its activation function to convert the raw output values into a probability distribution over possible classes. This allows the system to assign probabilities to each class, making it suitable for multi-class classification problems where the input data must be categorized into one of several predefined classes. The softmax function ensures that the output values are normalized, summing to 1, which facilitates interpretation as probabilities. This approach is particularly useful in applications like image recognition, natural language processing, and other domains requiring probabilistic classification. The system may be implemented in hardware or software and can be trained using labeled data to optimize the weights of the neural network for accurate classification.
14. The apparatus of claim 8 , wherein the one or more computer readable media further include instructions that when executed cause the apparatus to convert each of the words included in the document into a one-hot vector and generate the embedding vector for each of the words by a product of the embedding matrix and the one-hot vector for each of the words.
This invention relates to natural language processing (NLP) and text analysis, specifically improving document representation through word embeddings. The problem addressed is the need for efficient and accurate conversion of text into numerical representations that capture semantic meaning, which is crucial for tasks like document classification, search, and machine learning. The apparatus includes a processing system and one or more computer-readable media storing instructions. The system processes a document by converting each word into a one-hot vector, a binary vector indicating the word's presence in a predefined vocabulary. Each one-hot vector is then multiplied by an embedding matrix to generate an embedding vector for the word. The embedding matrix is a learned representation that maps words to dense, low-dimensional vectors, preserving semantic relationships. This process transforms the entire document into a sequence of embedding vectors, enabling further analysis or machine learning tasks. The invention improves upon traditional methods by leveraging matrix multiplication for efficient embedding generation, reducing computational overhead while maintaining semantic accuracy. This approach is particularly useful in applications requiring real-time text processing or large-scale document analysis. The system may also include additional components for preprocessing, such as tokenization or normalization, to enhance the quality of the embeddings. The overall goal is to provide a scalable and accurate method for converting text into meaningful numerical representations.
Unknown
January 21, 2020
Browse 5M+ US patents with plain-English claim translations and AI-generated analysis.